The AI data market, weekly · 19 Oct 2026

A $1 billion budget for company records, and prices nobody agrees on

Buyers of company records got bigger and louder this month: micro1 pledged $1 billion, Handshake built a sales floor and at least six more buyers are hiring to source data. The prices actually paid are far below the bands they advertise, the terms are rarely published, and Washington has started to notice.

Who is buying

  • micro1 committed to spend $1 billion over the next 12 months acquiring and licensing enterprise operational data, with published tiers of more than $100,000, more than $500,000 and more than $1 million per partnership (9 October). Its only measured pace so far: more than 50 data transactions in 45 days, as it told the Spirit Airlines court in September. MoreSource: micro1
  • Handshake AI staffed an outbound sales floor to license company workflow data, code and support records, advertising $100,000 to $4 million or more per deal. A 15 October job post says it pays about $60 million a month to more than 30,000 people. MoreSource: Handshake AI job posts
  • New buyer-side roles in the last ten weeks: Toloka hired a founding data partnerships lead to sign businesses that have never sold data (factories, warehouses, retail, services), Thinking Machines Lab opened a data partnerships role, SpaceXAI is building an RL environments team from real-world work data, and Microsoft AI and AfterQuery are hiring to acquire data. MoreSource: The companies' own job posts
  • Snorkel AI, which raised $350 million at a $3.5 billion valuation in September, opened a function to source and procure private data and says in the post that the economics of this market are still being invented.Source: Snorkel AI job post

What it actually sells for

  • MIT Technology Review followed one small closed startup: about $25,000 for its code base and about $15,000 pending for its Google, Notion, Figma and Slack workspace data, sold through SimpleClosure, which has handled over 150 such deals; 47% of its wind-downs now include a code or data sale. MoreSource: MIT Technology Review, 13 October
  • In the same report, Surge AI's environments lead says buyers quote wildly different prices for the same archive, because each values it on its own research thesis about how many useful training tasks it yields. One price from one buyer says little about what the data is worth. MoreSource: MIT Technology Review, 13 October
  • The one sale with public bids: Google won the August auction of Spirit Airlines' records (about 100 million emails, 500 million Teams messages and 176,000 employee records) at $10 million; micro1 then offered $12.5 million, 25% more, after the auction closed. The court has not yet approved the sale. MoreSource: Fortune, 14 September

Terms to read twice

  • Columbia Journalism Review published a draft Handshake contract offering a freelance journalist $1,000 per case of reporting notes for a non-exclusive, royalty-free, perpetual, irrevocable, transferable and sublicensable licence. MoreSource: Columbia Journalism Review, 15 October
  • SimpleClosure's terms give the buyer of a closed startup's assets all right, title and interest, to use them without further consent from or compensation to the company selling. Google's Content Offer Pilot, now expanding, goes the other way: holders keep their IP and may sell the same content elsewhere. MoreSource: SimpleClosure terms; Google Content Offer Pilot

Law and politics

  • More than 110 members of Congress, led by Senator Elizabeth Warren and Representative Steven Horsford, wrote to Google and Spirit on 8 October objecting to the sale of the airline's workplace records as training data: the first national political attention on company records. MoreSource: Yahoo Finance
  • The Third Circuit affirmed on 29 September that ROSS's training on Westlaw headnotes was not fair use, the first appeals-court ruling on the question, and found that the market for licensing such material as AI training data is rapidly developing. Weeks earlier the Department of Justice filed a statement backing fair use for training in another case. MoreSource: US Court of Appeals for the Third Circuit
  • The Anthropic authors' settlement, approved on 20 July, pays about $3,000 per work, four times the statutory minimum for willful infringement; first payments are due by 15 November. It settles how the books were acquired, not a licence, but it is the clearest per-work number the market has. MoreSource: Bartz v. Anthropic, final approval order

Elsewhere in the market

  • Figure says it will spend over $1 billion on data and compute in the next 12 months and launched Index, a paid first-person video programme, after vendors could not meet its bar; it has paid creators $15 million so far. MoreSource: Figure
  • Cloudflare added Pay Per Use beside Pay Per Crawl on 30 September: a buyer makes an offer, the publisher accepts or counters, and payment follows reported use. Wiley's AI licensing revenue fell to $14 million in its latest quarter, from $29 million a year earlier. MoreSource: Cloudflare; Wiley's SEC filing

The AI data market, weekly

Issues go out on Mondays: the prices, deals, buyers and rulings that moved in the market for AI training data, each with its source.

Free. Issues go out on Mondays, and every one has a link to leave. Subscribing means you accept our terms and privacy policy.

A $1 billion budget for company records, and prices nobody agrees on | AIData.Broker